Qwen3.8-27B 262K Context Memory Analysis

Analysis of the Qwen3.8-27B model card confirms a KV cache cost of 64 KiB per token. This implies that using the native 262,144 (262K) context requires 16 GiB of additional memory beyond the model weights.

Key technical features and hardware requirements are as follows:

• Hybrid Architecture: Of the 64 layers, 16 use Full Attention, while 48 use DeltaNet. DeltaNet layers maintain a fixed recurrent state (approx. 72 MiB), so memory usage does not increase with context length.
• Memory Requirements and Hardware Availability:
  • 16GB GPU: Almost no headroom for additional context when loading model weights.
  • 24GB GPU: Supports approx. 100K context with f16 KV, and approx. 180K context with q8 KV.
  • 32GB GPU: Can accommodate the full 262K context with q8 KV.
  • 80GB GPU: Allows using FP8 weights and full context while retaining headroom for Batching.

The model's hybrid DeltaNet layout provides approximately 4x cache savings compared to Full Attention, serving as a key factor in resolving memory bottlenecks.